Information Retrieval System for the Automatic Generation of Scientific Abstracts with a Pre-trained Model

This article presents a web system for retrieving scientific articles, with a module for automatic generation of abstracts using the pre-trained GPT-2 model, which takes the title of a given scientific article as input. To improve the results of text generation, a fine-tuning was made to GPT-2 by re-training the model, with 6102 scientific articles retrieved with Europe PMC’s RESTful API and using a set of keywords associated with computer science topics. The information retrieval system uses two similarity metrics, Jaccard and Cosine, to create a ranking of results according to the input of the query. These two similarity metrics were also used to measure the divergence between the original abstract and the one generated with the GPT-2 model. For the process of functional validation of the system, experiments were carried out with the keywords indexed in the list of lexemes for tuning the model. Two scientific articles were taken for each word. In addition, keywords outside the list were used, and 30 random experiments were carried out. The results showed a mean total similarity of 0.1064 and 0.2426 for the Jaccard and Cosine metrics. Although the results show a low level of similarity, it is essential to clarify that the original abstract of the article considered ground truth is a subjective text proposed by the authors. Therefore, this does not imply a system malfunction. Still, it shows that the methodology proposed in this article allows for the generation of coherent text with a small text string as input.